Anthropic’s New Usage Policy Bans Abusing Claude

Anthropic has published its yearly rewrite of the rules for using Claude. The update, posted on 8 October, takes effect on 12 November and mostly clarifies rules that already existed, the company says, while adding examples for the longer, more autonomous work Claude now does and responding to misuse it has seen in influence operations, weapons development and surveillance. One addition has drawn most of the attention: a ban on abusing the model itself.

A rule against cruelty to Claude

The new Usage Policy prohibits “sustained and needless abusive or cruel behavior toward our models.” Anthropic says it applies only in extreme cases where users repeatedly act cruelly toward Claude with no discernible purpose, and not to ordinary frustration, pushback, dark creative themes, or model testing and research. It follows a step the company took in August, letting Claude end rare conversations with persistently abusive users, and Anthropic says that ability remains the main way the rule will be enforced. In other words, the practical effect of being abusive is still that Claude hangs up; the policy now also says the behaviour is not allowed.

The addition lands in a wider argument about how to treat AI systems. TechCrunch notes that it follows reporting by The New York Times on Anthropic’s outreach to religious scholars, in some cases to make the case that Claude could be conscious or have a soul. The policy text itself makes no such claim.

Elections and fake accounts

Rules that were spread across sections on elections, fraud, privacy and disinformation now sit together under “Do Not Engage in Deceptive Campaigns or Artificial Activity.” It covers political and commercial work alike: hiding who is behind a message, amplifying content through fake accounts or posts, and building tools or infrastructure for influence operations. Anthropic points to its September threat intelligence report for the patterns behind the change.

The election section is renamed “Do Not Undermine Democratic Processes” and is aimed at deceiving voters and disrupting elections: false information about candidates or voting, impersonating candidates or officials, and suppressing turnout. At the same time, Anthropic has removed its blanket ban on personalised vote and campaign targeting, so ordinary civic and campaign work is allowed. Deceptive targeting and misuse of voters’ personal data stay prohibited under other sections.

Surveillance, weapons and machines that move

The surveillance and law-enforcement rules have been rewritten. Claude may not be used to track people without consent, either in real time or by analysing data already collected, to decide or recommend whom to investigate, arrest or charge, or to build or improve surveillance tools. Consented tracking such as fraud monitoring, content moderation, journalism and legal research remain allowed, and Anthropic says the rewrite clarifies its enforcement rather than changing it. The weapons rules now name software and components that make weapons work, and the arming of drones and other autonomous vehicles, which the company describes as formalising what it already enforced.

For high-risk uses that touch health, legal rights, money, livelihoods or essential services, the existing requirements stay: a qualified human must be able to review and change Claude’s recommendations, and the people affected must be told AI was involved. New is a rule for hardware. When Claude controls equipment that takes autonomous physical actions and could injure someone, a qualified operator must be able to watch it and stop it, and the equipment must hold a safe state if Claude is disconnected, in line with Anthropic’s Model Hardware Standard. The update also spells out who the supported regions rule shuts out: people physically in unsupported regions, companies incorporated or headquartered there, and companies majority-owned or controlled from them.

The abuse clause is the only part of the update about how people treat the model rather than what they use it for. The rest reads like a policy written for Claude’s work leaving the chat window, into fake-account networks, voters’ inboxes and machines that move, and spelling out who is answerable when it gets there.